File size: 5,627 Bytes
1f0ae0a
138b689
 
6999303
 
 
 
138b689
 
 
 
6999303
 
 
 
 
 
138b689
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6999303
 
 
 
 
 
138b689
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6999303
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
import os
import gradio as gr
import torch
import pydub
import torchaudio
from torchaudio.sox_effects import apply_effects_tensor
import numpy as np
from transformers import AutoFeatureExtractor, AutoModelForAudioXVector

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

def load_audio(file_name):
    audio = pydub.AudioSegment.from_file(file_name)
    arr = np.array(audio.get_array_of_samples(), dtype=np.float32)
    arr = arr / (1 << (8 * audio.sample_width - 1))
    return arr.astype(np.float32), audio.frame_rate

STYLE = """
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/css/bootstrap.min.css" integrity="sha256-YvdLHPgkqJ8DVUxjjnGVlMMJtNimJ6dYkowFFvp4kKs=" crossorigin="anonymous">
"""
OUTPUT_OK = (
    STYLE
    + """
    <div class="container">
        <div class="row"><h1 style="text-align: center">The speakers are</h1></div>
        <div class="row"><h1 class="display-1 text-success" style="text-align: center">{:.1f}%</h1></div>
        <div class="row"><h1 style="text-align: center">similar</h1></div>
        <div class="row"><h1 class="text-success" style="text-align: center">Welcome, human!</h1></div>
        <div class="row"><small style="text-align: center">(You must get at least 85% to be considered the same person)</small><div class="row">
    </div>
"""
)
OUTPUT_FAIL = (
    STYLE
    + """
    <div class="container">
        <div class="row"><h1 style="text-align: center">The speakers are</h1></div>
        <div class="row"><h1 class="display-1 text-danger" style="text-align: center">{:.1f}%</h1></div>
        <div class="row"><h1 style="text-align: center">similar</h1></div>
        <div class="row"><h1 class="text-danger" style="text-align: center">You shall not pass!</h1></div>
        <div class="row"><small style="text-align: center">(You must get at least 85% to be considered the same person)</small><div class="row">
    </div>
"""
)

EFFECTS = [
    ["remix", "-"],
    ["channels", "1"],
    ["rate", "16000"],
    ["gain", "-1.0"],
    ["silence", "1", "0.1", "0.1%", "-1", "0.1", "0.1%"],
    ["trim", "0", "10"],
]

THRESHOLD = 0.85

model_name = "microsoft/wavlm-base-plus-sv"
feature_extractor = AutoFeatureExtractor.from_pretrained(model_name)
model = AutoModelForAudioXVector.from_pretrained(model_name).to(device)
cosine_sim = torch.nn.CosineSimilarity(dim=-1)


def similarity_fn(path1, path2):
    if not (path1 and path2):
        return '<b style="color:red">ERROR: Please record audio for *both* speakers!</b>'
    
    wav1, sr1 = load_audio(path1)
    print(wav1, wav1.shape, wav1.dtype)
    wav1, _ = apply_effects_tensor(torch.tensor(wav1).unsqueeze(0), sr1, EFFECTS)
    wav2, sr2 = load_audio(path2)
    wav2, _ = apply_effects_tensor(torch.tensor(wav2).unsqueeze(0), sr2, EFFECTS)
    print(wav1.shape, wav2.shape)

    input1 = feature_extractor(wav1.squeeze(0), return_tensors="pt", sampling_rate=16000).input_values.to(device)
    input2 = feature_extractor(wav2.squeeze(0), return_tensors="pt", sampling_rate=16000).input_values.to(device)

    with torch.no_grad():
        emb1 = model(input1).embeddings
        emb2 = model(input2).embeddings
    emb1 = torch.nn.functional.normalize(emb1, dim=-1).cpu()
    emb2 = torch.nn.functional.normalize(emb2, dim=-1).cpu()
    similarity = cosine_sim(emb1, emb2).numpy()[0]

    if similarity >= THRESHOLD:
        output = OUTPUT_OK.format(similarity * 100)
    else:
        output = OUTPUT_FAIL.format(similarity * 100)

    return output


inputs = [
    gr.inputs.Audio(source="microphone", type="filepath", optional=True, label="Speaker #1"),
    gr.inputs.Audio(source="microphone", type="filepath", optional=True, label="Speaker #2"),
]
output = gr.outputs.HTML(label="")


description = (
    "This demo will compare two speech samples and determine if they are from the same speaker. "
    "Try it with your own voice!"
)
article = (
    "<p style='text-align: center'>"
    "<a href='https://huggingface.co/microsoft/wavlm-base-plus-sv' target='_blank'>🎙️ Learn more about WavLM</a> | "
    "<a href='https://arxiv.org/abs/2110.13900' target='_blank'>📚 WavLM paper</a> | "
    "<a href='https://www.danielpovey.com/files/2018_icassp_xvectors.pdf' target='_blank'>📚 X-Vector paper</a>"
    "</p>"
)
examples = [
    ["samples/cate_blanch.mp3", "samples/cate_blanch_2.mp3"],
    ["samples/cate_blanch.mp3", "samples/cate_blanch_3.mp3"],
    ["samples/cate_blanch_2.mp3", "samples/cate_blanch_3.mp3"],
    ["samples/heath_ledger.mp3", "samples/heath_ledger_3.mp3"],
    ["samples/russel_crowe.mp3", "samples/russel_crowe_2.mp3"],
    ["samples/cate_blanch.mp3", "samples/kirsten_dunst.wav"],
    ["samples/russel_crowe.mp3", "samples/kirsten_dunst.wav"],
    ["samples/russel_crowe_2.mp3", "samples/kirsten_dunst.wav"],
    ["samples/heath_ledger.mp3", "samples/denzel_washington.mp3"],
    ["samples/leonardo_dicaprio.mp3", "samples/russel_crowe.mp3"],
    ["samples/leonardo_dicaprio.mp3", "samples/russel_crowe_2.mp3"],
    ["samples/naomi_watts.mp3", "samples/denzel_washington.mp3"],
    ["samples/naomi_watts.mp3", "samples/leonardo_dicaprio.mp3"],
    ["samples/naomi_watts.mp3", "samples/cate_blanch_2.mp3"],
    ["samples/naomi_watts.mp3", "samples/kirsten_dunst.wav"],
]

interface = gr.Interface(
    fn=similarity_fn,
    inputs=inputs,
    outputs=output,
    title="Voice Authentication with WavLM + X-Vectors",
    description=description,
    article=article,
    layout="horizontal",
    theme="huggingface",
    allow_flagging=False,
    live=False,
    examples=examples,
)
interface.launch(enable_queue=True)